Every change to a score, a criterion, or the way the Machine Intelligence Score is calculated, with the date it took effect and the reason for it.
Scores on this site are recomputed when the rubric behind them is corrected, so a figure noted here one month may not match the one published the next. Every such change is listed below rather than made quietly.
Nothing on this page changes a score. It records changes already made, and the methodology explains how the current scores are calculated.
Two changes to the same criterion, made together, and worth stating plainly because of how they interact. First, MachineMetrics was recorded as having native vibration sensing and does not. Their own engineering post says the data behind their anomaly detection "does not contain: 1. Vibration 2. Power 3. Temperature 4. Any other data that would normally have to be collected by external sensors", and sells that as an advantage. Their hardware catalogue offers a generic industrial IoT sensor line that can carry vibration, but no vibration product of their own, unlike AssetWatch, Augury, Petasense or MOVUS. Third-party is the correct value. Second, third-party sensing now scores 7 rather than 5. It is the only middle anchor in this rubric that means the same capability sourced elsewhere, rather than a lesser version of it: a third-party accelerometer can produce better data than a vendor's own, and what a buyer loses is having to source, mount and integrate it. Halving the score overstated that when the value below, 0, means the platform cannot do it at all. The interaction is the part worth recording. The MachineMetrics correction on its own would have moved this site's own platform into first place. The scoring change is what keeps MachineMetrics first, at 81 against Sense on 80. Both changes are defensible on their own evidence and they were made in the same commit, so a reader can judge the pair rather than see only the half that suits us.
The six dimension weights were previously calculated from how much each dimension's criteria overlapped statistically. They are now set from shop owner interviews: sensing and predictive maintenance 20 each, and OEE tracking, commercial terms, integration depth and deployment 15 each. Most composites and most ranks change. MachineMetrics leads.
Sense was recorded as tracking OEE only partially, because nothing on sensemfg.com mentioned OEE and an audit of every citation on this site had demoted the entry for exactly that reason. The vendor has now published it ("Controller data adds the context for complete OEE tracking"), so the fact is citable and the score moves from 5 to 10. The capability is conditional: the same page says controller access is available for qualifying customers and never required. This rubric scores whether a platform can produce automated OEE, and 25 platforms are scored native on the same basis. This is the first-party platform on this site, and the change was made when the claim appeared on a public page rather than when it was asserted.
This criterion scored whether a buyer was offered a choice of capital treatment, so a perpetual licence sold alongside a subscription took full marks and everything else that was not a hardware-plus-subscription split took three out of ten. That put Vorne and FactoryWiz, who sell the software outright with no recurring fee at all, on the same score as a platform where the buyer rents everything and owns nothing when they stop paying. It now scores what a buyer ends up holding: software owned outright, or offered as both perpetual and subscription, scores 10; hardware bought with the software rented scores 6; nothing owned scores 3. Vorne, FactoryWiz and FANUC MT-LINKi rise. Sense is unaffected at 6.
This site briefly recorded Sense as having adapter-level controller access rather than native, on the grounds that its own page describes reading the machine controller as optional, over MTConnect, for qualifying customers. That was wrong, and it was wrong in the direction of being hard on ourselves rather than easy. Every other platform here that reaches a controller through a published protocol is scored native: MachineMetrics through FOCAS, MTConnect and OPC-UA, FANUC MT-LINKi through OPC UA, and twenty more. What this rubric calls adapters is an intermediary standing between the controller and the platform, such as the Petasense transmitter or the data historians that Prometheus reads. Sense reads the controller directly over MTConnect, so native is the consistent answer and the score returns to where it was. One question the rubric does not currently ask is whether a capability offered as an option to some customers should score the same as one that is how the product works by default. That distinction would affect several platforms, not only this one, and belongs in a rubric change rather than in a single score.
Each platform page now shows the figure its Price score is actually computed from: the vendor's published rate normalised to dollars per unit per year, so a per-machine subscription, a per-interface minimum and a perpetual licence spread over three years can be read against one another. Twelve of the 43 publish enough to compute it. For the other 31 the page now says so plainly, because for them the score turns on whether a buyer can see a number at all rather than on how large it is. The bands are terciles of the platforms that publish a rate, not lines drawn by hand, and the page says that too. This figure has been recorded against every platform since pricing was first sourced and cited, and had never been shown to anyone, so the Price score was a number with no visible basis, which is the one thing this site is not supposed to publish.
Each of the twelve criteria on a platform page is now a link to the specific page the score was read from, with the provenance and the date it was retrieved. That evidence was recorded for all 43 platforms and shown for none of them, which asked a reader to take twelve numbers on trust. Publishing it exposed five scores that were wrong, three of them too generous. TeepTrak was credited with dedicated power monitoring, when their current sensor detects machine cycles and none of their seven modules is an energy product, and with installation measured in minutes, when the ten to fifteen minutes covers mounting the sensor and their own page title offers to connect any machine in one hour. TeepTrak falls from 70 to 65 and from sixth to eighth. Petasense was published as bypassing the machine controller, when their transmitter documentation describes extracting PLC data over Modbus TCP through an Ethernet port; Petasense rises from 58 to 63 and from twentieth to eleventh. Scytec was recorded as vendor-installed, when their own installation page documents downloading and installing the software yourself, with consulting offered separately; Scytec rises from 55 to 60. RedZone and Fabrico were both recorded as charging for installation, when RedZone's pricing page includes guided deployment and on-site coaching with every deployment and Fabrico publishes free onboarding on every paid tier; RedZone rises from 40 to 43 and Fabrico from 61 to 65. Four of the five corrections favour a competitor and none favours us. How long installation takes and who performs it are now recorded as two facts rather than one. They were a single entry with a single citation feeding two separate scores, so where a vendor published one and not the other the entry had to be marked to the weaker of the two, and real evidence went unrecorded. Ndustrial states it goes live in weeks without saying who does the work; Scytec documents a self-serve install without giving a duration. Both are now recorded properly. Splitting them raises the number of recorded facts from 387 to 430, and the number resting on inference from 48 to 54, because a fact previously counted as half of another is now counted on its own. Further research brought that figure down to 50, and vibration monitoring is now evidenced for all 43 platforms with no inferences left in it. A wider audit of whether each cited page actually states the fact it is cited for then pushed the number back up to 98, which is the point of counting it. All 278 cited pages were opened, including the five vendors whose sites answer an automated request with a challenge page and had to be rendered in a browser. Around fifty facts were published as vendor-confirmed while resting on pages that never mention them. A FANUC product page cited for nine separate facts states one. Sight Machine's homepage is cited for eight and states two. Others pointed at general articles about a topic rather than the vendor's own answer, at software-listing aggregators, at a third-party manual host, at a vendor's blog post comparing two of its competitors, and in six cases at a page that had quietly started returning a 404. None of those facts is necessarily wrong. Each is now marked as what it always was: read from published material rather than stated by the vendor. Six further scores were confirmed against vendor material rather than left as inferences: TeepTrak states it provides APIs, Samsara describes fault codes and usage-based scheduling rather than predictive models, Amper states onboarding and training are included, C3 AI documentation enumerates its reliability capabilities without OEE tracking or energy metering, and MOVUS describes external sensing rather than a controller connection. IoTFlows publishes 'Starting at $119 /mo. + $149 one-time hardware cost' on its product page, which this site had recorded as publishing no figure at all; that alone moves its price score from 0 to 7. The same page states there is no PLC integration and that no knowledge of MTConnect, OPC UA or Modbus is needed, so its controller access is bypassed rather than adapted, moving that score from 5 to 0. IoTFlows rises from 73 to 75 and from fourth to third. FactoryWiz's published perpetual licence was also found to have risen from $2,195 to $2,310 per machine since it was last checked, and the recorded figure now matches. A wider audit of whether each cited page actually states the fact it is cited for then corrected several more. UpKeep publishes a power and electrical sensor line tracking current, voltage and resistance, so its power monitoring rises from partial to dedicated and UpKeep from 60 to 64. JITbase states setup takes less than 30 minutes, which is faster than the hours we had recorded; JITbase rises from 73 to 75. MOVUS integrates with PLC, SCADA and DCS through a line interface unit rather than bypassing the controller, and rises from 66 to 71. Guidewheel's installation charge came from an aggregator rather than the vendor and is now recorded as an inference. Several facts were demoted because the page cited for them never mentioned them: C3 AI's controller access and installation agency, Harmoni's predictive maintenance, and FANUC's power monitoring, all of which cited marketing pages that discuss none of those things. FANUC's ownership and contract term were correct but unsupported, and now cite the FANUC page stating MT-LINKi is a one-time software purchase. All told, 22 composites and 23 of the 43 ranks move, and the dimension weights were re-measured four times as the underlying facts changed. One citation was repaired: Ndustrial's installation evidence pointed at a page returning a 404, and a check of all 275 distinct citations found no other dead link.
What installation adds to the bill is now scored. Nineteen platforms include it: Tractian, Evocon and MachineMetrics state there is no setup charge, AssetWatch bundles on-site installation of up to 200 sensors into its monthly fee, and Mingo folds hardware, licences and onboarding into one first-year price. Five publish a figure, and they vary by two orders of magnitude: Sense charges $70 per unit, Scytec $1,000 to $2,500 to activate an account, UpKeep $500 to $5,000 for implementation packages, eMaint $2,000 to $15,000, and Guidewheel $5,000 to $20,000. Six charge without publishing a rate, including IBM, AVEVA, SKF, Emerson and Parker, all of which sell installation or commissioning as a separate service. Figures are compared at a five-machine shop so a per-unit fee and a per-project engagement can be read against each other. Charging without publishing a figure scores the same as charging a large one, because a buyer given no number has to budget for the worst. Twenty-three of the forty-three entries are inferences rather than vendor statements, the most of any criterion here, and that count is published on the methodology page.
These were one criterion and answered two different purchasing questions badly. Installation speed now scores duration alone: ten platforms install in minutes, nine in hours, ten in days, fourteen in weeks or longer. Self-install capability scores whether a buyer can do it without a technician on site: nineteen can, twenty-four cannot. Separating them means a vendor fitting sensors in twenty minutes is recorded as fast without being called self-service, and a three-week guided rollout is no longer indistinguishable from a same-visit one. The two move together across the platforms reviewed, correlating at 0.74, so the weighting gives the pair roughly the influence of one and a quarter criteria rather than two. Criteria now number eleven.
Every one of the ten scoring criteria is now calculated from facts recorded against each platform, each citing the page it came from. No score is assigned by hand. Machine age compatibility was removed rather than derived: its scale offered only "any machine regardless of age" or "modern CNC only", and every platform reviewed reaches the first, so the criterion could not distinguish between them. Criteria now number ten. The methodology page publishes how many recorded facts are vendor statements and how many are inferences, criterion by criterion.
Installation is now recorded as two separate facts: who performs it, and how long it takes. The published scale had conflated them, since "minutes self-install" describes a duration while "requires vendor" describes who does the work, so platforms landed differently depending on which half their own material happened to mention. Ten platforms are self-installed in minutes, nine self-installed over hours, ten installed by the vendor or an integrator within days, and fourteen take weeks or longer. Both facts now appear on each platform page, so a vendor fitting sensors in twenty minutes reads differently from a three-week guided rollout even though the scale treats both as not self-service. The evidence for this criterion remains weaker than for any other here, since most of it is vendors describing their own deployment times.
Three outcomes: a current or power sensor is part of the product, electrical condition is inferred from something else, or there is no electrical sensing. Twelve platforms sense power directly through a clamp, transducer or three-phase meter. Seven infer it indirectly, mostly from magnetic flux around a motor. Twenty-four have none. Thirty-six scores changed. Parker rises furthest: its SensoNODE Gold line includes current and power sensors, which this site had not reflected. SKF falls, because QuickCollect measures velocity, acceleration and temperature only.
Nearly every platform here markets the phrase "predictive maintenance". The scale asks whether a trained model is behind it, whether an alert fires when a reading crosses a threshold, or whether there is no predictive capability. Nineteen use trained models, nineteen use thresholds, five have none. Thirty-five scores changed. Tulip states it has no native predictive maintenance and no automated anomaly detection. Fabrico states its AI modules are not yet available. FANUC MT-LINKi and TeepTrak rise to full marks on anomaly scoring against a learned baseline.
Three outcomes: the platform speaks a controller protocol, it reaches machine data another way, or it does not touch the controller. Twenty speak a protocol such as MTConnect, FANUC FOCAS, OPC UA, Modbus or HART. Eleven use discrete inputs, a relay contact or an upstream historian. Twelve do not touch the controller, which is a design choice for sensor-first platforms and scores zero here. Thirty-eight scores changed. Tractian states its sensors work without needing PLC access, so it scores zero rather than 6. eMaint, Samsara, IBM Maximo and Emerson rise to full marks on published protocol support.
Three outcomes: OEE calculated automatically from machine signals, OEE that is partial or needs manual entry, or no production monitoring. Twenty-four calculate it automatically, nine partially, ten not at all. Thirty-eight scores changed. Limble does not connect to production machines and does not monitor availability, performance or quality. Prometheus Group requires a separate OEE platform. Augury, MOVUS, Petasense and Emerson are machine-health tools that improve OEE without measuring it. AVEVA rises, detecting stops automatically from live equipment data.
Three outcomes: the vendor supplies its own vibration sensors, vibration arrives from third-party hardware, or there is no vibration capability. Thirteen ship their own sensors, thirteen take the data from someone else’s, seventeen have none. Twenty scores changed. The enterprise and CMMS platforms take vibration from hardware they do not make: Maximo over MIF and MQTT, PI System through the SETPOINT gateway, eMaint from Fluke 3563 and 3561 sensors, Evocon through Arpedon MAINTNODE.
Three outcomes: a public API reference, an API with no usable documentation, or none. Fifteen publish a genuine reference. Twenty-six state they have an API but publish nothing a developer could build against, or return a sign-in page. Two have no API. Thirty-six scores changed. Vorne held a 4 while offering an API whose documentation its support team sends on request. Guidewheel held an 8 on a single sentence of marketing copy. UpKeep and Limble held 8s while publishing full public references.
Both criteria were described as based on reviews from G2 and Capterra. Those reviews do not exist for most platforms here: UpKeep has roughly 2,400 across the two and Limble roughly 1,400, but MachineMetrics has four, several have single figures and many have none. Forty-three scores rested on an evidence base covering about a dozen of them, and neither criterion had a defined scale, so no value could be checked against a rule. Both are removed.
Price was the last score assigned by hand and the only one with no defined scale, so no value could be wrong. Its published logic names two facts rather than a judgement, and the score is now calculated from both: whether a buyer can find the price, taken from the citation each platform already carries, and the published figure expressed per machine per year with hardware spread over three years. The boundaries between low, mid and high are the thirds of the twelve platforms that publish a figure. Twenty-seven publish no figure and score zero. The top score went from one platform to four: UpKeep, Scytec, FactoryWiz and JITbase. Slytek publishes a price list starting at $79 per machine per month, which this site had described as unpublished.
The weights are calculated from how much a dimension's criteria overlap, and that overlap is measured from the scores. One of the six figures had not been re-measured since the criteria underneath it were rescored, so the published weights described an older version of the data. Twenty-three of 43 composites were wrong by up to two points. All six figures are now measured from the scores as published, and the build stops if any of them drifts from the data again. No check like that existed before, which is why this went unnoticed for as long as it did. Commercial Flexibility now takes 25.8 per cent of the composite, up from 17.1 per cent. Its three surviving criteria measure close to independent things, once the redundant ones were removed and contract terms stopped being assigned by hand, and the formula gives a dimension influence in proportion to how much it independently measures. Every other dimension gives up a little to make room. The correlation figures quoted on the methodology page are now calculated when the site is built, because the hand-written ones went stale without anyone noticing.
Seat policies and order minimums are no longer scored. Neither could be answered from what vendors publish: only 11 of 43 state how many users a plan includes, and only 12 say whether a minimum order applies, so 43 published scores rested on facts that existed for about a quarter of them. Order minimums also tracked price almost exactly, at a correlation of 0.94, so the two were largely one measurement counted twice. Both are still reported on each platform page where the vendor publishes them, and marked as not published where nobody does. Contract terms were kept, because most vendors do state a term. That score is now calculated from the term recorded for each platform instead of assigned by hand, which changed 30 of the 43. The published rubric defines three outcomes, no contract, annual and multi-year, while the scores in use ran across 2, 3, 4, 6 and 7. Eight of the terms are read from what a vendor publishes about billing rather than from a stated term, and the methodology page gives that count. Two corrections came out of the same research, both against us. AssetWatch is an annual commitment and not the no-contract arrangement we published, because terminating after the first year costs half the remaining fees. Slytek publishes a price list this site described as unpublished.
This criterion asks whether a buyer gets a choice between owning the system and renting it. It had stopped answering: 34 of 43 platforms shared a single value, and two scores contradicted the pricing published on their own review pages. Every platform’s acquisition model has now been established from the vendor’s own terms, cited and dated: whether the hardware is bought outright or bundled into the subscription, and whether the software is available perpetually, by subscription, or both. The score is now calculated from those two facts rather than assigned, so it can no longer disagree with the evidence beside it. Eight scores changed. Inductive Automation, IBM Maximo and eMaint were recorded as offering one route to purchase when each offers both. Scytec was recorded as offering both when it publishes only a subscription. MachineMetrics, Samsara, Nanoprecise and IoTFlows sell their hardware outright while renting the software, which had not been reflected.
The methodology described a 1–10 scale while eight of the criteria define 0 as a legal outcome and 22 published scores were 0. The scale is 0–10, and the pages now say so. Separately, the scoring table on this page is now generated from the same definition the build validates every score against, so the rubric a reader is shown and the rule the scores are held to cannot drift apart. That check found that a majority of scores sit on values their own rubric does not define: a 7, for example, on a criterion that offers only 10, 5 and 0. Correcting those needs each one re-examined against evidence rather than snapped to the nearest allowed value, so the count is now recorded and the build refuses to let it grow. No score changed in this release.
Each platform’s pricing was previously stated as fact with no source and no date. All 43 have now been checked against the vendor’s own pages, and each carries a citation saying where the figure came from and when it was last verified. Twenty-three were wrong or misleading and have been rewritten. The largest errors were in both directions: Vorne was published at about $1,995 per line when the vendor’s cheapest listed model is $4,490; Scytec, Evocon, JITbase, Guidewheel and Tractian were described as quote-only when each publishes a full price list. AssetWatch’s $199 was a one-time 30-day trial fee, not a monthly rate. Where a figure could not be traced to any source, including one of our own, it has been removed rather than left standing.
The rankings page sorted on a second composite, overallScore, which followed no published rule and matched the criteria mean for 8 of 43 platforms. It has been removed. The Machine Intelligence Score is now the only score the site publishes, and it is what every ranking and sort reads.
Four changes to what gets measured. Commercial terms were five criteria measuring close to one thing, so dimensions are now weighted by how much their criteria actually vary independently rather than by how many there are. Predictive Maintenance and OEE Tracking were averaged together despite moving in opposite directions, and are now scored separately. Edge/Cloud was removed: it published 43 scores backed by six documented facts. Ownership Model was rescored to measure whether a buyer gets a choice of capital treatment, rather than preferring one purchasing model over another.
The site published a scoring formula its own numbers did not follow: only 7 of 43 platforms matched it. All 43 are now computed from the weights published on the methodology page, and the build fails if they ever drift again. 38 of the 43 published scores changed as a result, by as much as 16 points up and 12 down. No platform’s underlying criteria ratings were edited; only the arithmetic that turns them into a composite.